Re: lazy detoasting

From: Chapman Flack <chap(at)anastigmatix(dot)net>
To: Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us>
Cc: Peter Eisentraut <peter(dot)eisentraut(at)2ndquadrant(dot)com>, Amit Kapila <amit(dot)kapila16(at)gmail(dot)com>, Robert Haas <robertmhaas(at)gmail(dot)com>, Andrew Gierth <andrew(at)tao11(dot)riddles(dot)org(dot)uk>, PostgreSQL Hackers <pgsql-hackers(at)postgresql(dot)org>
Subject: Re: lazy detoasting
Date: 2018-04-11 18:38:46
Views: Raw Message | Whole Thread | Download mbox | Resend email
Lists: pgsql-hackers

On 04/11/2018 01:55 PM, Tom Lane wrote:
> Chapman Flack <chap(at)anastigmatix(dot)net> writes:
>> Well, the devilsAdvocate() function would stash the object
>> in a static, then try to look at it some time in a later call
>> in the same transaction.
> If you're worried about that, you should also worry about what happens
> if the function uses the static variable in some later transaction.
> The spec grants you license to throw an error, but it still needs to
> be a clean error (not something about "can't find toast value", IMO).

There's precedent for that kind of thing in PL/Java already ... objects
that Java considers alive as long as some code holds a reference
to them, but proxy for things in PG that may only have function-call
lifetime or cursor-row lifetime, etc. If they are closed by Java code
(or the Java GC finds them unreachable) first, they have to remember
to release their PG stuff; if the PG stuff goes first, they have to
update themselves to throw a suitable "you've kept me past my sell-by
date" exception if the Java code tries to use them again.

Thomas implemented most of those things ages ago; this is the first
I've added myself, with a little adjustment of technique because his
were for lifetimes shorter than transaction. I'm using the
TopTransactionResourceOwner to learn when the transaction is finished.
Resource owners have been around as long as any PG version PL/Java
supports, so that seems ok.

> Can you detect that the value is being stored in a long-lived variable
> and detoast at that point?

Not easily, I don't think. The question resembles "is this object
still reachable, or unreachable, now as I exit this function call?"
or at some other specific time. The Java garbage collector eventually
learns what's become unreachable, but it doesn't promise *when*.

But let me return to the earlier idea for a moment: are you saying
that it might *not* be sufficient to find an applicable snapshot at
the time of constructing the object, and register that snapshot
on TopTransactionResourceOwner?

It would obviously not be kept around any longer than the transaction,
and would be released earlier whenever the Java code reads/closes it,
or lets go of the last reference and GC finds it. In all the typical
cases I can imagine, it would be registered very briefly, with the
object being constructed/read/freed in quick succession.


In response to


Browse pgsql-hackers by date

  From Date Subject
Next Message Pavan Deolasee 2018-04-11 18:43:18 Re: Bugs in TOAST handling, OID assignment and redo recovery
Previous Message Tom Lane 2018-04-11 18:34:46 Re: WARNING in parallel index creation.